Modeling and Generating Wi-Fi Traffic with Stochastic Models and Neural Networks

This thesis studies synthetic generation of IEEE 802.11 MAC traffic at the level of frame subtype, inter-frame space (IFS) and duration. We collect real-world traffic captures in a shielded measurement chamber and preprocess them into cleaned time series. We propose two approaches for generating such time series, aiming to cap-ture strong inter-feature dependencies and to follow the rules dictated by the IEEE 802.11 MAC protocol. First, we propose a simple, interpretable stochastic model, NGramKDE, which samples subtype sequences via an N-gram language model and generates IFS and duration from subtype-conditioned logKDE estimators. Second, we design PacketGAN, a conditional WGAN-GP architecture with several generator and discriminator variants. We compare PacketGAN variants against NGramKDE and the DoppelGANger baseline using quantitative metrics (including Discrimina-tive Score and Signature MMD) and qualitative analyses (including t-SNE and manual inspection). We find that the best-performing PacketGAN variants and NGramKDE generate high-quality samples and achieve strong performance across the evaluation methods considered.